20% off on annual plans this month
code SMILE2020% off on annual plans this month
code SMILE20One panoramic film, three separate clinical questions before you book an extraction: is the tooth impacted, which way is it angled, and how close does the root actually sit to the nerve. AI answers each with a different confidence level — here's what the published data says about all three.
What this article covers
How accurately AI detects impacted third molars on a standard OPG, how well it classifies angulation using Winter's and Pell & Gregory systems, what the data shows on AI's ability to flag proximity to the inferior alveolar nerve — and why that last one still deserves a CBCT before anyone picks up a handpiece.
Third molar extraction is, by most counts, the single most common procedure an oral surgeon performs. It's also one of the few dental procedures where the pre-op read genuinely changes the treatment plan — not just the diagnosis. A mesioangular impaction sitting clear of the canal is a routine chairside extraction. The same tooth horizontally impacted with its root crossing the inferior alveolar canal is a referral, a CBCT, and a very different conversation with the patient about nerve injury risk.
That's three separate reads stacked into one film:
Each of those three has a different amount of published evidence behind it. Lumping them into one “can AI read a wisdom tooth X-ray” question is where most of the confusion — and most of the marketing overreach — comes from.
Well-Validated Today
Genuinely Harder
Where AI Helps Most
Two classification systems, doing two different jobs
Winter's classification describes the angle of the impacted tooth relative to the second molar — mesioangular, distoangular, vertical, horizontal. Pell & Gregory describes depth and available space (Class I–III, position A–C). Surgeons use both together to gauge difficulty before they touch a handpiece, and it's exactly this combination that recent detection models have been trained to reproduce.
Start with detection, because it's the strongest number on this page. A 2025 model built on VGG16 for detection and ResNet50 for classification — trained and tested on 1,100 panoramic radiographs containing 1,200 impacted mandibular third molars — hit 93.51% detection accuracy. Not a lab curiosity. An oral radiologist validated every annotation first.
93.5%
Detection accuracy
Impacted mandibular third molars, OPG
J. Oral Med. Oral Surg., 2025
0.98
mAP@50, angulation classification
Winter's + Pell & Gregory, YOLOv11
Scientific Reports, 2025
72.3% vs 52.7–69.6%
True nerve-contact accuracy
AI vs. six OMFS specialists
Scientific Reports, 2023
That third card is the one worth actually reading twice. Six oral and maxillofacial surgery specialists — not students, not general dentists — were asked to determine true root-to-canal contact from a panoramic film alone. Their accuracy ranged from 52.68% to 69.64% — a coin flip would get you most of the way to the bottom of that range. The AI model, trained against CBCT-verified ground truth, came in at 72.32% — and on a related sub-task, judging bucco-lingual position of the nerve relative to the root, the gap widened further: AI at 80.65% against a specialist range of just 32.26% to 48.39%.
Three reads, three very different accuracy ceilings
How to read thisDetection
Is a tooth impacted, and where? Mature, well-studied, over 93% accuracy across multiple architectures.
Angulation
Winter's / Pell & Gregory classification. Also strong — precision above 0.95 in recent object-detection models.
Nerve proximity
Contact and bucco-lingual position. Improving fast, beats unaided human reads on OPG, but still a screening signal — not a surgical verdict.
A panoramic radiograph flattens a three-dimensional jaw into a two-dimensional strip. Somewhere in that flattening, the actual spatial gap between a root tip and the inferior alveolar canal gets lost — a root can appear to overlap the canal on film while sitting a full millimeter buccal or lingual to it in real anatomy. That's not a training-data gap AI can simply be fed more of. It's a physical limitation of the imaging modality itself.
Which is exactly why the numbers above matter. AI isn't overcoming that limitation — nobody has. What it's doing is reading the same ambiguous 2D signal more consistently than a tired surgeon on their eleventh OPG of the day, and doing it against a CBCT-anchored ground truth most human readers never get to calibrate against in daily practice.
Dentist plus AI still beats either one alone
One study measured average precision on nerve-relationship detection three ways: dentists working alone (76.45%), an AI model working alone (83.02%), and dentists using the AI as a second opinion (88.06%). The pattern shows up everywhere this gets tested rigorously — assisted humans outperform either the unaided clinician or standalone AI. Nobody comes out ahead by removing the other from the workflow.
"Ground truth" itself is contested. Different studies verify true nerve contact against different references — some against CBCT, some against surgical findings, some against senior-radiologist consensus. Accuracy numbers across studies (52% to 89%+) aren't always measuring the exact same thing, which is worth remembering before treating any single figure as gospel.
Canal wall continuity is a weaker read than contact alone. One 2022 model that scored 0.860 accuracy on contact detection dropped to 0.766 on the more surgically relevant question of whether the canal's cortical wall stayed intact — a meaningfully harder sub-task that current models haven't fully cracked.
A "low risk" flag on OPG isn't a green light. Standard surgical protocol routes panoramic radiographs showing classic radiographic signs of proximity — darkening of the root, interruption of the canal's white line, deflection of the root — into CBCT confirmation before finalizing a surgical plan. AI screening doesn't change that referral pathway; it changes how fast you get to the decision to refer.
Angulation classification is mature; the nerve question isn't there yet. Don't let a strong 98% mAP on Winter's classification create false confidence about a 72–89% accuracy range on a much harder, much higher-stakes anatomical question.
No AI output carries surgical sign-off. The oral surgeon's license and clinical judgment — not a confidence score — is what stands behind the decision to extract, refer, or scan further.
| Task | AI on OPG | Clinician on OPG alone | Where the difference comes from |
|---|---|---|---|
| Detecting the impacted tooth | Strong (93.5%) | Strong | Mature, well-represented task for both |
| Angulation classification | Strong (0.98 mAP) | Strong | Deterministic geometry, easy to label and train on |
| True nerve contact | Moderate (72–89%) | Weak (53–70%) | 2D projection loses real depth; AI trained on CBCT-anchored data |
| Canal wall continuity | Early-stage (~77%) | Not reliably tested | Subtler radiographic sign, less training data available |
| Decision to extract or refer for CBCT | Not applicable | Strong | Clinical judgment and licensure, not pattern recognition |
Medecro's AI X-Ray Analyzer runs exactly this triage on OPG and RVG — impaction detection, angulation classification, and a proximity-to-IAN risk flag, with confidence scores and one-click override, inside the workflow you already use. It's built to speed up which third molars need a closer look, not to replace the CBCT or the surgeon's final call on the ones that do.
Medecro AI X-Ray Analyzer
Confidence-scored detection built for third molar workup, with one-click override, inside your existing clinic workflow. It sorts which cases need a closer look — it doesn't replace the CBCT or the surgeon's call on them.
Book a Demo — See It LiveYes — this is one of the more mature applications on this page. Recent object-detection models trained on Winter's classification report precision above 0.95 and mAP@50 around 0.98 when tested against radiologist-labeled panoramic datasets.
Sources & references
